1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-22 02:33:06 +01:00

X86: fixup (V)PMADDWD detection

Fix some bugs (missing checks).
Add constant support.
This commit is contained in:
Nekotekina 2021-05-17 08:18:19 +03:00
parent f7d625e31a
commit c9fceef173

View File

@ -43536,9 +43536,11 @@ static SDValue combineMulToPMADDWD(SDNode *N, SelectionDAG &DAG,
// with zero extends, which should qualify for the optimization. // with zero extends, which should qualify for the optimization.
// Otherwise just fallback to zero-extension check. // Otherwise just fallback to zero-extension check.
if (isa<ConstantSDNode>(N0.getOperand(1).getOperand(0)) && if (isa<ConstantSDNode>(N0.getOperand(1).getOperand(0)) &&
N0.getOperand(1).getConstantOperandVal(0) == 16 &&
isa<ConstantSDNode>(N1.getOperand(1).getOperand(0)) && isa<ConstantSDNode>(N1.getOperand(1).getOperand(0)) &&
N1.getOperand(1).getConstantOperandVal(0) == 16) { N0.getOperand(1).getConstantOperandVal(0) == 16 &&
N1.getOperand(1).getConstantOperandVal(0) == 16 &&
DAG.isSplatValue(N0.getOperand(1)) &&
DAG.isSplatValue(N1.getOperand(1))) {
// Nullify mask to pass the following check // Nullify mask to pass the following check
Mask17 = 0; Mask17 = 0;
N0 = DAG.getNode(ISD::SRL, N0.getNode(), VT, N0.getOperand(0), N0 = DAG.getNode(ISD::SRL, N0.getNode(), VT, N0.getOperand(0),
@ -43547,8 +43549,20 @@ static SDValue combineMulToPMADDWD(SDNode *N, SelectionDAG &DAG,
N1.getOperand(1)); N1.getOperand(1));
} }
} }
if (!DAG.MaskedValueIsZero(N1, Mask17) ||
!DAG.MaskedValueIsZero(N0, Mask17)) if (!!Mask17 && N0.getOpcode() == ISD::SRA) {
if (isa<ConstantSDNode>(N0.getOperand(1).getOperand(0)) &&
DAG.ComputeNumSignBits(N1) >= 17 &&
DAG.isSplatValue(N0.getOperand(1)) &&
N0.getOperand(1).getConstantOperandVal(0) == 16) {
Mask17 = 0;
N0 = DAG.getNode(ISD::SRL, N0.getNode(), VT, N0.getOperand(0),
N0.getOperand(1));
}
}
if (!!Mask17 && (!DAG.MaskedValueIsZero(N1, Mask17) ||
!DAG.MaskedValueIsZero(N0, Mask17)))
return SDValue(); return SDValue();
// Use SplitOpsAndApply to handle AVX splitting. // Use SplitOpsAndApply to handle AVX splitting.